Improve fireball detector recall with stdpixel decontamination

[alextudorica]

Summary

The fireball detector misses events when a bright fireball inflates stdpixel
along its trail within the 256-frame FF block. The threshold formula
(avepixel + k1*stdpixel + j1) exceeds 255, gets clipped to 255 (uint8), and
the fireball becomes invisible at the very pixels where it is brightest (87%
of fireballs saturate at maxpixel=255).

The line finder (find3DLines) is not the problem -- it has zero failures
when given a clean point cloud. The bug is entirely upstream in the threshold
stage.

Changes

VideoExtraction.py -- stdpixel decontamination (main fix)

Before thresholdAndSubsample, replace stdpixel at contaminated pixels with
the background median:

1. Background reference: std_ref = median(stdpixel) where maxpixel < P90
2. Contamination mask: maxpixel >= P90 AND stdpixel > 3 * std_ref
3. Replace: stdpixel[contaminated] = max(1, round(std_ref))

Pure numpy preprocessing. Zero-copy fast path when no contamination is present.

Grouping3Dcy.pyx -- threshold clip 254

Clip threshold to 254 (not 255) so near-saturated pixels (maxpixel=255) can
pass. Without this, maxpixel > 255 is always false for uint8.

Grouping3D.py -- FOV-aware velocity filter

Replace the hardcoded if total > 2 in findCoefficients with
fireball_max_ang_vel (deg/s) converted to px/frame via FOV and resolution.
Addresses the existing TODO: "this limit should be read from config file and
calculated for FOV". Falls back to the legacy total > 2 when no config is
passed.

.config / ConfigReader.py -- fireball_max_ang_vel

New config parameter (default 60.0 deg/s). Wider than the meteor detector's
ang_vel_max (35 deg/s) because the fireball detector is recall-first.

Benchmark results on 129-event reference dataset

1755 real FF files, 743 stations (639 detected by stock, 104 missed).

FR-positive recall (686 FFs where stock produced an FR file):

Gate	Count
DETECTED	679 (99.0%)
velocity	5
min_frames	2
threshold	0
test_points	0
line_finder	0

Zero threshold failures on real data confirms the decontamination works. The 5
velocity failures also fail with the legacy total > 2 filter (not regressions
from the FOV-aware change -- the decontamination changes the point cloud
geometry which shifts line slopes past the velocity cap in both cases).

Missed station rescue (104 stations where stock produced no FR):

	Stock	With decontamination
Stations rescued	76/104 (73.1%)	77/104 (74.0%)

Overall station-level recall:

	Stock	This PR
Stations detected	639/743 (86.0%)	~716/743 (96.4%)

The 716 is an upper bound -- the 7 FF-level failures may cause a small number
of station-level losses (floor: 709/743 = 95.4%).

27 stations remain unreachable: 13 genuinely faint (threshold), 10 too brief
(min_frames < 4), 2 velocity, 2 overexposed (white_avg).